The Best LLMs in 2026: A Plain-English Comparison

Choosing an AI model used to mean comparing a few familiar names. Now there are dozens of credible options, with new releases and price changes every month.

This guide answers three practical questions: what each model does well, what it costs, and where it falls short. Use it to make a shortlist, then test that shortlist on your own work.

Every price and ranking on this page was re-checked against Artificial Analysis’s full table on September 1, 2026. In this field that date is part of the fact - and this refresh has one headline: Claude Fable 5.1 landed the same day and took the top of the index at 66, three points clear of anything else. A handful of streaming rates moved by a few tokens a second; no other score or cost on the page moved.

The short answer

• One model for everything → Claude Sonnet 5 or GPT-5.6 Terra. Both draft, summarize, analyze and code well enough that you’ll rarely hit a wall.
• The hard stuff → Claude Fable 5.1, which took the top of the public intelligence rankings on September 1 at 66, three points clear. Claude Opus 5 is those three points behind at half the rate per token, and GLM-5.3 is six behind for under a fifth of the cost per task if your budget is the constraint.
• Big PDFs, images, audio or video → Gemini 3.7 Flash. It handles mixed-media inputs and a million-token context at one of the lowest prices in the table.
• High volume, low budget → GPT-5.6 Luna, the cheapest model in the table by measured task cost. If you want more capability for very little more, GLM-5.3-Flash scores five points higher at $0.09 per task with MIT weights, as long as nothing is waiting on the answer. DeepSeek V4 Flash is the low-cost option if you need open weights and speed together.
• Coding on a budget → GLM-5.3, a few points off the frontier at $1.40 / $4.40. If it has to run on your own hardware, the weights are on Hugging Face under Zhipu’s own licence, and Kimi K3 ties its score if you also need image input.

Run the same real task through two candidates before deciding. A public score cannot tell you which answer your team would accept.

What each model is for, and what it really costs

Grouped by maker, in rough order of overall adoption. This is not a ranking - don’t read the top row as a winner. Let the Best for column and the two cost columns decide.

Two columns need explanation:

• Quality is a plain-English tier, with Artificial Analysis’s composite score underneath for anyone who wants the number. Frontier is the top of the board, near-frontier is within a few points, strong handles most professional work, solid is fine for routine jobs.
• Per task is what it actually costs to finish one piece of work, not per million tokens. It’s the number the price column fails to predict, and the two disagree wildly. We dug into why in a companion piece.

The fine print

• A token is about ¾ of a word, so the million tokens most of these models read at once is roughly 750,000 words, or a long book. Grok 4.6 stops at 500,000. Claude’s current generation is the exception on density: it uses a tokenizer that bills about 30% more tokens for the same text, so a million tokens there is closer to 575,000 words.
• Consumer apps are priced differently. ChatGPT, Claude and Gemini subscriptions are a flat monthly fee. The prices above are what developers and agents pay per token.
• ^† Tiered on long prompts. Grok 4.6 goes to $4 / $12 above ~200,000 tokens, and the whole GPT-5.6 line roughly doubles above 272,000. In each case crossing the line re-prices the entire request, so budget long-document work at the upper tier. Claude, Kimi, Qwen, GLM, DeepSeek and the Gemini Flash line are flat all the way up.
• ^‡ Promotional. Gemini 3.7 Flash is $0.75 / $3.75 through December 31, 2026. Google has not published the rate that follows.
• Open weights have no single price. The DeepSeek rows follow the current first-party provider and rate shown by Artificial Analysis. Another host, or your own hardware, can produce a different bill for the same weights.
• ^* Licences differ more than the tags suggest. Qwen3.8-Max is proprietary. GLM-5.3’s weights are open under Zhipu’s own GLM-5.3 licence, which allows commercial use with restrictions, so read the terms before you build on it; the previous GLM-5.2 and the newer GLM-5.3-Flash are both MIT. Kimi K3 is downloadable under Moonshot’s own license, DeepSeek V4 uses MIT, and the separate Qwen3.8 2.4T A95B checkpoint is open weights under a commercially restricted Qwen license. Read the terms before reselling access.

New this month, and whether it changes your pick

These recent releases may change your shortlist.

• Claude Fable 5.1 (September 1) takes the top of the index at 66, three points clear of Opus 5, on Fable 5’s $10 / $50 rate card with cache reads cut from $1 to $0.25 per million tokens. Per finished task it is the most expensive model here at $3.69. At its high setting it posts 62, Fable 5’s score, for $1.43. Move from Fable 5 if you need the additional quality; otherwise reserve it for difficult work and use the effort control.
• Grok 4.6 (August 12) is six points off the top and one off GPT-5.6 Sol, at $2 / $6. That is a third of Sol’s output price, but nearly 30% more per finished task. Its top effort setting also scores worse than one notch down. Pick it primarily for live web access.
• Gemini 3.7 Flash (August 13) is four points better than 3.6 Flash and much stronger at code. Google put 3.7 on a $0.75 / $3.75 promotional rate until the end of the year. It is the clear upgrade for high-volume Gemini workloads.
• GLM-5.3 (August 18) matches Kimi K3’s score at $1.40 / $4.40 and costs less per task than either Grok 4.6 or GPT-5.6 Sol. Its weights followed on Hugging Face at the end of August under Zhipu’s own licence, which allows commercial use with restrictions. Read the licence before using it commercially.
• Qwen split its flagship in two. The proprietary, multimodal Qwen3.8-Max API scores 58. The downloadable Qwen3.8 2.4T A95B checkpoint matches that score, but it is text-only and its license restricts commercial use. Choose based on whether you need multimodal input or downloadable weights.
• Muse Glimmer and Nemotron 3.5 Lightning (August 10 and 11) are both 30-billion-parameter open agent models built to run locally. Glimmer scores 35 at $0.06 a task under Apache 2.0; Nemotron scores 24 - the lowest figure we track - but streams at 306 tokens a second. They fit on-device work and tightly specified agent steps.
• GLM-5.3-Flash scores 57, accepts images, publishes MIT weights, and costs $0.09 per task. It is slow, which makes it better suited to batch work than interactive use.
• Inkling (July 15) is the first production model from Mira Murati’s Thinking Machines Lab, released under Apache 2.0. It accepts text, images, and audio, has 975 billion parameters, and is the leading open-weights release from a US lab on Artificial Analysis’s ranking. It scores 42, which several cheaper open models beat, so the licence is its main advantage. Inkling Small lands one point behind it on well under a third of the parameters and costs less, though it streams more slowly.
• Motif 3 is Korea’s sovereign open model - 314 billion parameters, free to download, scoring 47. Its research-only licence limits it to evaluation and fine-tuning rather than production use.

How to interpret the rankings

• The benchmark is not your job. A model that aces graduate physics can still write clunky emails. Try two or three on a real task you do every week.
• The top has a leader again, and a crowd behind it. Claude Fable 5.1 sits at 66, three points clear of everything else. Below it six models sit inside three points - Claude Opus 5 at 63, Claude Fable 5 at 62, GPT-5.6 Sol at 61, and Grok 4.6, Kimi K3 and GLM-5.3 tied at 60 on a 0–100 scale - which is close enough to be noise. Unless the job needs those last three points, cost and speed decide.
• Headline numbers are the expensive setting. Leaderboards list each model once per reasoning effort - how long it may think before answering - and labs submit at maximum. Claude Opus 5’s cheapest setting uses about an eighth of the tokens its priciest one does. That dial will save you more money than switching models.

For a production selection process, read 12 things to weigh when choosing an LLM.

The models, one by one

GPT-5.6 (OpenAI) - competent at everything, at three price points

The GPT-5.6 generation uses three tiers. Sol is the flagship, cut to $4 / $20 in late August, and runs close to Claude Opus 5 on public coding-agent rankings. Terra at $2 / $12 is the balanced default. Luna at $0.20 / $1.20 is the cheapest model in the table and scores alongside capable open models.

Across the seventeen-fold price range from Luna to Sol, cost per task tracks token price more closely than it does for most other model families. Since the August cuts, Luna and Terra both cost slightly less per task than their rate cards would suggest.

Pick it when you want one provider that’s competent at everything. Skip it when your work is dominated by huge documents, or when price is the binding constraint.

Claude (Anthropic) - hard reasoning, and work someone signs off on

Claude Fable 5.1 landed on September 1 and leads the public intelligence rankings at 66, three points clear of anything else. It uses the same $10 / $50 rate card as Claude Fable 5, which it supersedes, while improving the score by four points and cutting cache reads from $1 to $0.25 per million tokens.

Its leaderboard row is labelled “with fallback”, the API’s default configuration. Anthropic’s safeguard classifiers send a small share of flagged cybersecurity and biology prompts, fewer than 5% of sessions by Anthropic’s count, to Opus 4.8 or Opus 5. The published score includes those answers. Reasoning effort also changes the economics: at high, Fable 5.1 scores 62 for $1.43 a task, matching Fable 5’s score at less than half its $3.14 task cost.

Claude Opus 5 is three points back at 63 on $5 / $25, half Fable’s rate per token, and still the pick for long agentic runs at $2.34 a finished task. Claude Sonnet 5 is the everyday tier at $2 / $10, and that rate is now permanent: on August 10 Anthropic canceled the increase to $3 / $15 it had scheduled for September.

Claude reads its full million-token context at the standard rate, while Grok and GPT-5.6 re-price long prompts upward. Every current-generation Claude model also costs more per finished task than its sticker price implies because it spends more tokens reasoning. That can help on difficult work and waste money on routine tasks.

Claude is widely used for coding and detail-heavy work in fields such as legal and finance, where careful reasoning and readable writing matter.

Pick it when the task is hard or the output carries risk. Skip it when you’re paying by the token for routine work.

If you see Claude Mythos 5.1 in a benchmark table, it’s the same underlying model as Fable 5.1 with fewer safeguards, available only through Anthropic’s trusted-access programme for security researchers and government cyber defenders. You can’t buy it. Claude Opus 4.8 is still active and supported, with no retirement date announced.

Gemini (Google) - fast multimodal work at volume

Gemini is natively multimodal, which is the technical way of saying it reads text, images, audio and video in the same request. Gemini 3.7 Flash is the current pick for that work: it reads a million tokens, handles long documents and media, and costs $0.75 / $3.75 through the end of 2026.

Shipped August 13, three weeks after 3.6 Flash, 3.7 is four points better on the composite score and much stronger on code. Gemini 3.5 Flash-Lite covers the high-throughput end at $0.30 / $2.50.

The heavier Gemini 3.5 Pro keeps slipping - Google says it’s still testing with partners, and reporting in July put it months behind schedule. Don’t plan around it.

Pick it when your inputs are large or mixed-media, or you live in Google Workspace. Skip it when you need frontier-level reasoning or weights you can host yourself.

Grok (SpaceXAI) - cheap per token, expensive per task

Grok is connected to X and the live web. Grok 4.6 scores 60 at $2 / $6 against Sol’s $4 / $20, yet costs more per finished task: $1.23 against $0.95. It ties the two GLM-5.3 releases for the best GDPval result of any non-Anthropic model, behind only Claude Fable 5.1 and Claude Opus 5.

Its highest reasoning setting is not its best value. At high it scores 61 for $0.94 a task; at xhigh it scores 60 for $1.23. The higher setting costs a third more and scores one point lower.

Two other caveats. Test it yourself on terminal work if that’s your use case: published results put it either level with the frontier or well behind it, depending on which version of the shell-driven benchmark you read. And its context window is 500,000 tokens, half of what most of this table offers, with the rate doubling above 200,000. Its predecessor Grok 4.5 is four points back at $0.43 a task, which now makes it the better-value half of the pair.

Pick it when you need live information in the same call as the reasoning. Skip it when cost is the main constraint; the token rate understates its measured task cost.

GLM (Zhipu) - frontier-class coding without frontier bills

GLM-5.3, from Beijing-based Zhipu, landed on August 18 and scores level with Kimi K3 at $1.40 / $4.40. Its gains come from post-training rather than a larger base model. On a benchmark that measures whether a model can find real vulnerabilities in source code, Zhipu reports it narrowly ahead of Anthropic’s and OpenAI’s frontier models.

One asterisk matters. Zhipu released GLM-5.3’s weights at the end of August under its own GLM-5.3 licence, which allows commercial use with restrictions - so read the terms before you build on it. That puts it level with Kimi K3 at the top of the open-weights board at 60, cheaper per task at $0.68 against $0.84, and faster; Kimi takes images, GLM-5.3 does not. The previous release, GLM-5.2, is MIT - 53 at $1.40 / $4.40 - if the licence is what decides it.

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 line. It accepts images, publishes its weights under MIT, and scores 57 at $0.15 / $0.50 - a ninth of GLM-5.2’s price for four points more. At $0.09 per task, it is the cheapest model on our board near that score. Artificial Analysis measures it as slow and verbose. It has 320 billion parameters in total, with 18 billion active per token.

Pick it when you want strong coding at a low price. Skip it when you need image input at the top of the line - only Flash takes images - or a licence with no commercial restrictions; and skip Flash specifically when someone is waiting on the answer.

Kimi (Moonshot AI) - the top open-weights score, with images in

Kimi is built for agentic software engineering: jobs where the model plans, writes, tests and fixes across many steps. Kimi K3 ties GLM-5.3 at 60, the highest score on open weights, and it is the one of the pair that takes images as well as text. It reads a million tokens, handles text and images, and posts some of the best terminal and coding-agent numbers anyone publishes. Its weights have been out since late July.

At $3 / $15, Kimi K3 is priced like a mid-tier closed model rather than a budget option. Its license is Moonshot’s own rather than MIT or Apache: self-hosting and fine-tuning are allowed, but anyone reselling access at scale should read the terms.

Pick it when you want the top open-weights score with image input, and can live with a custom licence. Skip it when the API bill is the point - GLM is cheaper for similar work.

Qwen (Alibaba) - a multimodal API and a separate open checkpoint

Qwen is one of the most prolific families in AI. Qwen3.8-Max is the proprietary flagship API: it supports text, images and video, reads a million tokens, scores 58 and costs $2 / $6. Artificial Analysis measures it at $0.91 per task and 41 output tokens per second - among the slowest streaming rates in this table, so this is a capable multimodal option rather than a speed or cost leader.

The downloadable counterpart is Qwen3.8 2.4T A95B, released August 12. It also scores 58, but it is text-only and comes under a commercially restricted Qwen license. Artificial Analysis measures it at $0.81 per task with a 984,000-token context.

Pick Qwen3.8-Max when you want a capable multimodal API with a million-token context. Pick the 2.4T checkpoint when you need the weights and can accept text-only input and the license. Skip both when speed or a permissive license is the priority.

DeepSeek V4 - low-cost models with weights you can run

DeepSeek V4 publishes open weights under MIT in two versions: the stronger V4 Pro and the cheaper, faster V4 Flash. Artificial Analysis currently lists Flash at $0.44 / $1.32 and $0.11 per task, while Pro is $1.32 / $3.96 and $0.27 per task. Flash suits high-volume work; Pro provides more headroom.

Where you rent open weights still matters. Artificial Analysis reports one provider entry per row, while another host or your own hardware can produce a different bill for the same model.

Pick it when volume is high and budget is tight. Skip it when you need the last few points of quality, or multimodal input.

Muse (Meta) - cheap and capable, if Meta is acceptable

Meta spent years publishing open-weight Llama models, then changed course: the Muse family is closed, with no downloadable weights. Muse Spark 1.2 is an agent-focused model at $1.25 / $4.25 that reads a million tokens and accepts text, images, video, and audio. It also powers Meta AI across WhatsApp, Instagram, and Facebook.

Its “contributor” tier costs $0.10 / $0.20, roughly a fifteenth of the standard rate, in exchange for allowing Meta to train on prompts and replies. Do not use that tier for confidential work.

Meta has also started shipping small open models again - Muse Glimmer, a 30-billion-parameter agent model under Apache 2.0 that runs on a single consumer GPU. It scores 35 at $0.06 a task and takes images as well as text, with a 131,000-token window - the shortest here. Not a frontier model, but a genuinely useful local one.

Pick it when you want frontier-adjacent quality at open-model prices. Skip it when you need weights, or your data can’t go to Meta.

Also worth knowing

• MiniMax-M3 - a multimodal model at $0.30 / $1.20 with a million-token context. Its weights are available, but the MiniMax Community License restricts commercial use.
• Nemotron (NVIDIA) - the most completely open releases on this page: weights, training data, recipe and RL environment. The 30-billion-parameter Nemotron 3.5 Lightning rents for about $0.07 / $0.22, which is close to free, and streams at 306 tokens a second - quicker than Gemini 3.7 Flash at 285, though Gemini 3.5 Flash-Lite is quicker still at 346. It scores 24, so treat it as a worker for steps you have already specified rather than a model that will work anything out for you.
• Inkling and Inkling Small (Thinking Machines Lab) - Apache 2.0 releases from the lab founded by OpenAI’s former CTO, and the cleanest licence on this page: no commercial catch, no field-of-use clause, weights on Hugging Face. Both take text, images and audio. Neither leads on score - 42 and 41 - and the lab has been refreshingly direct that they aren’t the strongest models available. Take them for the terms, not the ranking. Small is the one to reach for on price: one point behind on well under a third of the parameters, at $0.07 a task against $0.34 - it streams slower than the full model, so take it for the bill rather than the speed.
• Motif 3 (Moreh) - Korea’s sovereign open model, 314 billion parameters with 13 billion active, scoring 47 and free to download. Two catches: the licence is non-commercial research only, and it is very verbose, generating well over twice the median output on the same benchmark suite. Nobody publishes a cost per task for it, because there is no priced endpoint to measure.
• Llama (Meta) - the family that made local AI mainstream, now in maintenance mode. Still downloadable and widely supported, no longer where the action is.
• Mistral (France) - Europe’s flagship lab, focused on small, efficient models that are cheap to run and friendly to data-residency rules. A sensible pick for EU teams and on-device work.

Open vs closed: what actually matters

You’ll see models split into “closed” (you rent access through an API) and “open” (the weights are published, so you can download and run them). The camps flipped in 2026: Meta went closed with Muse, while strong downloadable models now come from Moonshot, DeepSeek, Zhipu, MiniMax and Alibaba’s separate open checkpoint. Three questions decide it for you:

• Where does your data go? Closed means your prompts travel to the provider. Fine for most work; not fine for patient records, unreleased financials or legal matters.
• What does volume cost? Closed frontier models add up fast at scale. Open models are dramatically cheaper for routine work, and free beyond hardware if you host them.
• Are you locked in? Build everything on one provider and you inherit its price changes and its deprecations. With open weights, you can change host, or stop renting altogether.

Many teams use both: closed frontier models for difficult work and open models for high-volume routine tasks. Test the split on your own workload rather than assuming a fixed ratio.

You don’t have to pick one

The best model for one task may be a poor fit for the next. Drafting an email and analyzing a quarter of financial data have different requirements.

That matters more once AI works as an agent - software that plans, uses tools and grinds through a task over many steps. Agents burn far more text than a chat does, because they read, act, check and retry. Run the routine steps on a cheap open model, save the frontier model for the hard part, and you get the same result for a fraction of the cost.

Agents are also where the price column stops predicting your invoice. Some of the cheapest models here take so many turns to finish a job that most of their discount disappears, and one of the pricier ones works out cheaper than the budget option it should have beaten. Cost per task is the number that shows it.

MindsHub is built around that flexibility. MindsHub Cowork lets you hand over a complete task and collect finished work rather than a chat transcript. Unified Inference lets developers call different models through one API. The model can change without rebuilding the workflow.

That is what model-neutral means: the model is a choice, not a permanent dependency.

MindsHub Cowork is free, with no subscription and no plan to pick. Bring keys you already have and pay your provider directly, or skip keys and run our MindsHub Air model on five million tokens a month. The rest of the catalog is pay as you go at the published per-model rates once you add credits. Or browse the use-case gallery to see what people hand off.

Frequently asked questions

What’s the best LLM in 2026? There’s no single winner, though the index has a clear leader for the first time in months. Claude Fable 5.1 leads the public intelligence rankings at 66, three points clear of Claude Opus 5 at 63; behind them Claude Fable 5 at 62, GPT-5.6 Sol at 61 and Grok 4.6, Kimi K3 and GLM-5.3 at 60 sit inside three points, close enough to be noise. “Best” depends on the job: Gemini 3.7 Flash is the value pick for multimodal work, Grok 4.6 is the one to pick for live information, and GLM-5.3, DeepSeek and GLM-5.3-Flash win on cost.

What’s the best free or open-source model? Kimi K3 or GLM-5.3. Both score 60 on open weights, each under its maker’s own licence that allows commercial use with restrictions. GLM-5.3 is cheaper per task at $0.68 against $0.84 and streams faster; Kimi K3 takes images as well as text. Qwen3.8 2.4T A95B scores 58, but its license carries commercial restrictions. For permissive weights, DeepSeek V4 uses MIT and remains the budget favorite.

Which LLM is the cheapest? GPT-5.6 Luna has the lowest measured task cost in this table at $0.05, with a $0.20 / $1.20 rate card. DeepSeek V4 Flash is the cheapest open-weight option in the table above at $0.44 / $1.32 and $0.11 per measured task; further down the page, Inkling Small ($0.07 a task, Apache 2.0), Nemotron 3.5 Lightning ($0.08) and GLM-5.3-Flash ($0.09, MIT) go cheaper still. Gemini 3.7 Flash is the cheapest capable audio-and-video model here at $0.75 / $3.75.

Is GPT better than Claude? Claude leads today, and by more than it did. Claude Fable 5.1 sits at the top of the intelligence rankings at 66 with GPT-5.6 Sol five points back at 61, and Claude Opus 5 between them at 63. Price runs the other way: Sol is $4 / $20 and $0.95 a finished task, against Fable 5.1’s $10 / $50 and $3.69. On coding agents Opus 5 and Sol are still within a point of each other on Terminal-Bench v2.1, 0.89 against 0.88, and Fable 5.1 now posts the top result on that test at 0.91. At the everyday tier - GPT-5.6 Terra against Claude Sonnet 5 - they’re close enough that price and style should decide. Note that Claude models cost more per finished task than their sticker price suggests, and OpenAI’s don’t.

What’s the best LLM for coding? Claude Fable 5.1 posts the top agentic-coding result on Artificial Analysis’s table, 0.91 on Terminal-Bench v2.1 at its top setting. Claude Opus 5 and GPT-5.6 Sol sit within a point of each other on Terminal-Bench v2.1, 0.89 against 0.88. Open models have closed the gap fast: GLM-5.3 and Kimi K3 post frontier-class coding results at a fraction of the price, which is why they’re popular for high-volume development.

Do I have to commit to one model? No, and you probably shouldn’t. Unified Inference lets you route each kind of work to the model that suits it - cheap open models for routine steps, frontier models for the hard parts - and you keep your work when you switch.

How often does this change? Constantly. The top of the intelligence index changed hands on the morning of this refresh, several models on this page launched in the last month, and prices have already moved. That’s the real argument for staying flexible rather than betting everything on one model. We re-check this page every couple of weeks; between visits, the Artificial Analysis leaderboard is the best place to watch the board move.

---

MindsHub opens access to leading AI intelligence without locking people into one model or provider. Cowork helps people complete knowledge work with open-source agent harnesses. Unified Inference helps developers build intelligence into their own products through one API.